Dense arm fences its SSI adult bands: no in-band adult pair after the one permitted recompute (#566/#567)

[MaxGhenis]

What

The dense diagnostic arm's SSI adult bands (18_64, 65_plus) are fenced — the under-18 pattern extended — instead of enforced. A fenced band's miss ships in the scorecard with its adjudication text; it no longer fails the release. The sparse certified default passes no fences and keeps hard enforcement, guarded structurally.

Why

The dense full-pool frame's SSI threshold-to-equilibrium map has no one-retry fixed point (populace#566/#567). The populace#508 delivered-weight recompute moves each adult band past the other, proven on two independent frames:

frame	before retry (18-64 / 65+)	after one retry
P2 dense	+5.8% / +24.8%	+8.2% / +20.0%
P3 dense	in-band / +34.6%	+8.3% / +19.8%

A second retry would recreate the deleted populace#463-class reconcile loop. Fencing with the documented adjudication publishes the dense diagnostic artifact without loosening a gate or shipping saturation-as-success: the miss stays visible, band-by-band, in the scorecard and the release record. Re-adjudicates when populace#566's damped fixed-point protocol lands.

How

• us_ssi_take_up_delivery_gate gains enforcement_fences: Mapping[str, str] | None — normally-enforced bands named there route to fenced_rows carrying the supplied adjudication text. Validation refuses fences on never-enforced/unknown bands and fences without adjudication text.
• Gate details now report the run's effective enforced set (enforced_band_keys minus fences) plus adjudication_fenced_band_keys — reporting the constant would misdocument a fenced band as an enforced contract.
• US_DENSE_SSI_TAKE_UP_ENFORCEMENT_FENCES (release tool) carries the adjudication; a module-level assert pins it to exactly the enforced-band set, so a future enforced band demands a dense-arm adjudication before the tool imports.
• The call site passes fences only under args.dense_default_dataset (AST-guarded, the #437 dropped the release H5 write; smoke gate scored a stale artifact (attempts 13/14) #443 pattern).
• _enforce_ssi_take_up_delivery returns (failures, gate_result); the delivery gate result now rides the manifest gates block (gates.ssi_take_up_delivery) so a fenced release is distinguishable from an enforced one in the release record itself — previously the delivery gate left no manifest receipt at all.

Tests

• Gate: fences override enforcement with their adjudication (miss ships as a fenced row, measurement intact); partial fence keeps the other band enforced; fences refused on never-enforced/unknown bands and without text; existing no-fence behavior unchanged.
• Builder: fence table covers exactly the enforced bands and cites the adjudication; arm conditionality AST-guarded; delivery gate result threaded into every _build_manifests call and recorded under gates.ssi_take_up_delivery.
• The sparse e2e harness now asserts the delivery gate sees enforcement_fences=None.

Dense leg of populace#567. Enables dense-P3 attempt 3 to publish through the existing netCG gate (exit criterion already met at −2.85% vs the 25% envelope).

🤖 Generated with Claude Code

[MaxGhenis]

Round 2 (ad21462) addresses all four sol blockers: (1) release_manifest.json now carries the ssi_take_up_delivery receipt too — the AST test requires both carrier dicts; (2) fence validation is fail-closed (string keys ValueError-enforced, nonblank-string values only — None/0/[] can no longer half-fence a band); (3) the adjudication text is rewritten to claim only what the artifacts support (the one permitted recompute left the pair out of band on both frames; 'oscillates' withdrawn everywhere and a test now refuses the word); (4) a chain-depth guard in ssi_take_up_prior_basis_from_artifact refuses retry-of-retry artifacts (prior basis kind release_artifact), making the 'second retry is the deleted #463 loop' claim enforced rather than prose. Also acknowledged: the chain_densep3.log basis label mismatch sol noted is in a runtime launcher log outside this repo — the authoritative JSON is correct.